The moment came, as these moments often do, with Claude producing a site that was technically correct and visibly wrong. The layout held. The colors were defensible. But the spacing felt arbitrary in a way the maker could not unsee, and the proportions were slightly off in a way that no amount of "make it more like the reference" seemed to fix.
The maker of Taste Lab came away from repeated attempts with a conclusion that reframed the problem. The bottleneck was not the model. The bottleneck was the language being fed into it.
That hypothesis is now a product. Taste Lab launched this week on Product Hunt as a tool that extracts what its maker calls a website's design DNA, the colors, type, and spacing, and more importantly the reasoning behind each choice, and packages that reasoning as an artifact for an AI tool to read before it builds.
The framing the maker uses to explain it is blunt. "Tokens say what a design is. Taste adds the why: the trade-offs that explain each specific decision." That formulation is doing a lot of work. It claims that the missing ingredient in AI-assisted design replication is not better models or longer prompts, but explicit, site-specific reasoning that the model can be handed before it tries to render anything.
Whether that claim holds is the actual story. A single Product Hunt launch post is not enough evidence to settle it, and a reporter would normally wait for independent users, usage data, or a side-by-side test against existing style-transfer and design-token tools. The maker has built one. The maker has also framed the problem in a way that makes the product sound inevitable.
That framing deserves a closer look.
A central design choice in Taste Lab is a deletion rule. The system is built to throw out any extracted principle that could have been written without seeing the specific site being analyzed. The phrase "clean and modern" is the canonical example, a description that fits almost any site and therefore explains none of them. The maker's argument is that this kind of language is exactly what has been training AI tools to produce the "technically correct, clearly not it" results that triggered the project in the first place.
Positioned for the vibe-coding workflow, the extracted DNA is meant to be handed to a downstream AI tool, not read by a human designer. The pipeline assumes the next step is a build, not a design review. Product Hunt tags place Taste Lab in Design Tools, Artificial Intelligence, and the looser category of vibe coding, where a person describes a desired result in plain language and an agent produces the code.
The architectural idea is straightforward: if descriptive language is the bottleneck, then produce better descriptive language, formatted as something a tool can consume. The harder question is whether the bottleneck actually sits where the maker says it does. There are at least two alternatives the framing leaves unexamined.
One is that the failure mode the maker describes, with proportions off and spacing arbitrary, comes from the model's attention to reference images and color values rather than from the language in the prompt. A token-level description, no matter how careful, may not move a multimodal model away from a flat reading of the reference. In that case, extracting "the why" would not change the output, because the model is not reasoning about the why at all.
Another is that the gap between "technically close" and "it" is a judgment call that lives with the human, not the model. Designers often know a site is right or wrong before they can articulate why. If that is the case, then the most valuable artifact would not be the explicit reasoning Taste Lab extracts, but a tight feedback loop that lets a designer redirect the model in real time, in whatever vocabulary happens to work. Tools that already do this exist, and they do not require a "design DNA" intermediate.
The maker's claim is worth taking seriously because the frustration is real. Anyone who has asked a capable model to rebuild a known site has felt the "technically close, clearly not it" result. The open question is whether the fix is a new kind of intermediate artifact or simply a different relationship to the model.
What to watch is straightforward: does Taste Lab, or a tool like it, change the output of a downstream agent in measurable ways that older style-transfer and design-token systems do not. Until that question is answered by users, benchmarks, or comparison tests, "tokens say what a design is; taste adds the why" remains a hypothesis from one maker, a clean way to describe a problem, and a product built on top of that description.